Amey-Thakur Agent progressive disclosure
Sending every task to the strongest model with the fullest context is simple and expensive, and often worse, since a large context degrades attention. Progressive disclosure starts cheap and escalates on evidence, which usually handles most work at a fraction of the cost.
Method
- Start with the smallest viable context and model. Most tasks are routine, and routine tasks do not need the flagship model or the whole corpus.
- Define the escalation trigger explicitly. Low confidence, a failed verification, or an explicit cannot-answer. Escalating on vibes means escalating always.
- Let the agent say it lacks context. An agent that can request more is far more useful than one that guesses, and the request itself tells you what was missing.
- Escalate along one axis at a time. More context or a stronger model, so you learn which one mattered rather than paying for both.
- Cache what escalation revealed. If a class of task always escalates, promote it to start at the higher tier rather than paying the failed cheap attempt each time (see agent-specialist-router).
- Cap the ladder. After the top tier, the answer is a human, not another retry.
- Measure the escalation rate. Rising rates mean the cheap tier no longer fits the work, which is a signal to re-tune rather than to raise limits.
Boundaries
This optimises cost and latency; it adds a failed first attempt to every escalated task, so it loses on workloads that are uniformly hard. Confidence signals from models are unreliable, so triggers based on verification outcomes are sounder. Never let cost optimisation weaken a check on consequential work (see agent-human-checkpoint).